TY  - RPRT
TI  - Involuntary In-Context Learning: Exploiting Few-Shot Pattern Completion to Bypass Safety Alignment in GPT-5.4
AU  - Alex Polyakov
AU  - Daniel Kuznetsov
PY  - 2026
UR  - https://arxiv.org/abs/2604.19461
ID  - 2604.19461
ER  -